Yes, but not in the way we’d expect. When we adjust for the fixed effect of month, we get a negative correlation between lagged precipitation and incidence.
We calculate the distance from the centroid of each district to the South African border.
The distances can be seen in the below table.
Visually, we assess the hypothesis that the decline over time is steeper for those districts bordering South Africa through a series of charts.
As an alternative to the above, we can visualize the same data with the same aesthetics, but overlaying each year.
Could the proximity to South Africa partly explain the pre-MALTEM decline in incidence in Magude? We’ll examine this by modeling incidence as a linear function of district, distance, the interaction of year and distance, and month, in effect constructing a “counter-factual” for Magude that takes into account the secular decline in incidence associated with proximity to South Africa. For the model’s purpose, we take into account only those 4 years from 2012-06-01 through 2016-06-01 (so as to not include the effect of MALTEM). Our model formula predicts incidence as a function of the interactions of season/year and distance/year, with the fixed effect of month.
At first glance, ITN coverage appears associated with lower malaria incidence
Here’s what happens when we model cases using ITN, IRS, time since IRS campaign, seasonality, etc.
What are the mathematics of the method by which we assign weights to synthetic control districts? Are we concerned that so few districts get any weight?
Given that proximity to South Africa seems to have an interactive effect with the decline in malaria incidence, should we include this in our model? If so, how?
Are we currently excluding Manhiça from our synthetic controls? If so, why?
What are we doing with IRS / ITN?
How are we treating secular trends (ie, near linear decline in springtime malaria in Magude for the 4 years from 2012-2015)?